irspack — implicit recommenders for practitioners

irspack helps practitioners build, evaluate, and tune recommenders from implicit feedback such as clicks, views, saves, and purchases.

It is designed for the work that happens before a recommender reaches production: establishing strong baselines, comparing algorithms under the same evaluation protocol, and tuning promising candidates without rewriting the pipeline for every model.

• Practical recommenders including iALS, SLIM, item/user KNN, P3alpha, RP3beta, TopPop, and experimental feature-aware iALS
• Fast C++/Eigen implementations for core training and evaluation operations
• Consistent evaluation and Optuna-backed hyperparameter tuning
• Utilities for converting business IDs to sparse matrices and back
• Support for cold-start experiments with user and item side features

Read the documentation, see which recommender to try first, or start with your own interaction data.

Installation

irspack requires Python 3.9 or later. Install the published package with:

pip install irspack

Pre-built wheels are published for supported Linux, macOS, and Windows platforms. See Installing from source if a wheel is not available for your environment or if you want CPU-specific compiler optimizations.

Quickstart

irspack consumes a SciPy sparse matrix whose rows are users and whose columns are items. The values represent interaction strength; binary values are a good default for events such as clicks or purchases.

importnumpyasnpimportpandasaspdfromirspackimportIALSRecommender, df_to_sparseevents=pd.DataFrame(
{
"user_id": ["alice", "alice", "bob", "bob", "carol", "carol"],
"item_id": ["A", "B", "B", "C", "A", "D"],
}
)
# Rows and columns in X correspond to user_ids and item_ids, respectively.X, user_ids, item_ids=df_to_sparse(events, "user_id", "item_id")
model=IALSRecommender(X, n_components=8).learn()
# Recommend unseen items for Alice. Scores for already-seen items are -inf.alice_index=np.flatnonzero(user_ids=="alice")[0]
scores=model.get_score_remove_seen(np.array([alice_index]))[0]
top_items=item_ids[np.argsort(scores)[::-1][:2]]
print(top_items)

For an offline comparison, split interactions into training and validation matrices and evaluate every candidate with the same evaluator:

fromirspackimportEvaluator, IALSRecommender, TopPopRecommenderfromirspack.splitimportrowwise_train_test_splitX_train, X_validation=rowwise_train_test_split(
X, test_ratio=0.2, ceil_n_heldout=True, random_state=0
)
evaluator=Evaluator(X_validation, cutoff=3)
forrecommender_classin (TopPopRecommender, IALSRecommender):
recommender=recommender_class(X_train).learn()
print(recommender_class.__name__, evaluator.get_score(recommender)["ndcg"])

For timestamped events, prefer a temporal holdout over a random split. The using your own data guide shows the full workflow, including stable ID mappings and leakage-aware evaluation.

Which model should I try first?

Situation	Good starting point
Sanity check and popularity baseline	TopPopRecommender
Strong general-purpose collaborative filtering	IALSRecommender
Explainable item-to-item recommendations	CosineKNNRecommender
Sparse implicit-feedback data	RP3betaRecommender or SLIMRecommender
Cold-start users/items with side information	Feature-aware iALS

There is no universally best recommender. Start with a cheap baseline, compare a small set of candidates using a split that reflects the product scenario, then tune the winner. See the model selection guide for trade-offs and optional dependencies.

Hyperparameter tuning

Every tunable recommender exposes the same Optuna-backed interface:

best_params, trials=IALSRecommender.tune(
X_train,
evaluator,
n_trials=20,
random_seed=0,
)
best_model=IALSRecommender(X_train, **best_params).learn()

Iterative recommenders can use intermediate validation scores to stop unpromising trials early.

Optional recommenders

BPRFMRecommender wraps LightFM and requires a separate LightFM installation:

pip install lightfm

MultVAERecommender requires jax, jaxlib, dm-haiku, and optax. Follow the JAX installation guide if you need GPU support.

Installing from source

A source build requires a C++17 compiler. To compile using the instruction set available on the current machine:

CFLAGS="-march=native" pip install git+https://github.com/tohtsky/irspack.git

If installation fails, please open an issue and include your OS, Python version, CPU architecture, and the complete build error.

Development

This repository uses uv for reproducible local development:

uv sync
uv run pytest

Install documentation dependencies and build the site with:

uv sync --group docs
uv run sphinx-build -b html docs/source docs/_build/html

Use uv lock --upgrade only when intentionally updating dependencies. The lock file is committed so local development and CI use the same versions.